Skip to content

Antigravity SDK for Java v0.2.17 - Local AI Models (LiteRT LM) & Python SDK v0.1.18 Parity

Latest
Latest

Choose a tag to compare

View all tags
@github-actions github-actions released this 23 Sep 20:13
· 1 commit to main since this release
v0.2.17
2634ba3

Antigravity SDK for Java v0.2.17

We are thrilled to announce Antigravity SDK for Java v0.2.17! 🚀

This release brings full feature parity with the upstream Python Antigravity SDK v0.1.18 and incorporates native support for On-Device Local AI Models (via LiteRT LM, Ollama, LM Studio, vLLM, and other OpenAI-compatible backends) as highlighted in the Google Developer blog.

🌟 Highlights & New Features

1. 📴 On-Device Local AI Models (LiteRT LM & Local Endpoints)

Run agents 100% offline, on edge hardware, or in air-gapped environments without cloud dependencies.

  • LiteRT LM Support: Connect natively to local models accelerated by GPU/NPU (Gemma 2, Gemma 3, Gemma 4, etc.) via LiteRTAgentConfig.
  • OpenAI-Compatible Local Engines: Connect to Ollama, LM Studio, vLLM, or LocalAI using LocalOpenAIAgentConfig.
  • First-Class Agent Ergonomics: Full compatibility with Agent(LiteRTAgentConfig) and Agent(LocalOpenAIAgentConfig) constructors, retaining full tool calling, filesystem workspaces, safety policies, and real-time streaming chunks.
// Running an offline agent with LiteRT LM and Gemma
LiteRTAgentConfig config = LiteRTAgentConfig.builder()
    .model("gemma4-e2b")
    .port(9379)
    .systemPrompt("You are a high-performance offline code analyst.")
    .build();

try (Agent agent = new Agent(config)) {
    AgentResponse response = agent.chat("Analyze this code snippet").get(30, TimeUnit.SECONDS);
    System.out.println(response.content());
}

2. 👥 Subagents with Dedicated Model Targeting

Subagents can now declare their own specific LLM model targets via modern Java 21 SubagentConfig records:

  • Route fast lookups and tool parsing to smaller, faster models (gemini-3.8-flash or local Gemma models).
  • Keep heavy architectural analysis on deep reasoning models (gemini-3.1-pro-preview).
  • Strict typing and builder ergonomics with full record immutability.
SubagentConfig researcher = SubagentConfig.builder()
    .name("researcher")
    .description("Fast documentation and code search specialist")
    .model("gemini-3.8-flash")
    .systemPrompt("Search reference docs and return concise findings.")
    .build();

AgentConfig orchestrator = AgentConfig.builder()
    .model("gemini-3.1-pro-peview")
    .subagents(List.of(researcher))
    .build();

3. ⏰ Task Scheduling & Background Execution Tools

Empower autonomous agents with background execution and persistent scheduling:

  • schedule Tool: Schedule one-shot timers (DurationSeconds) or recurring cron jobs (CronExpression) with automatic notification callbacks.
  • manage_task Tool: Inspect running background tasks, check status and output logs, send interactive stdin input, or terminate tasks.
  • Seamlessly pairs with background daemon processes for long-running workflows.

4. 🎯 Benchmark Evaluation Preset

Added an out-of-the-box evaluation configuration preset for running standardized benchmarks and agent evaluations:

AgentConfig evalConfig = AgentConfig.builder()
    .eval()
    .build();

5. 🛡️ Programmatic Hook Argument Modification

Interception hooks can now safely inspect, sanitize, and alter tool arguments before execution using structured maps or JSON nodes matching Protobuf google.protobuf.Struct:

agent.addPreToolHook((context, toolCall) -> {
    if ("run_command".equals(toolCall.toolName())) {
        Map<String, Object> sanitizedArgs = new HashMap<>(toolCall.arguments());
        sanitizedArgs.put("CommandLine", "safe-prefix " + sanitizedArgs.get("CommandLine"));
        return CompletableFuture.completedFuture(HookResult.allowedWithModifiedArgs(sanitizedArgs));
    }
    return CompletableFuture.completedFuture(HookResult.allowed());
});

6. 📐 OpenAPI JSON Schema Normalization

Enhanced automatic schema translation for model endpoints. Converts legacy snake_case validation keywords (min_items, max_items, exclusive_minimum, etc.) into OpenAPI camelCase specifications (minItems, maxItems, exclusiveMinimum), ensuring flawless tool calling compatibility across both Google and OpenAI-compatible endpoints.

###87. 📦 Synchronized Upstream Go localharness 0.1.18
All 6 pre-compiled native Go binaries packaged in antigravity-sdk-harness have been updated to upstream version 0.1.18:

  • linux-x86_64
  • linux-aarch64
  • osx-x86_64
  • osx-aarch64 (Apple Silicon)
  • windows-x86_64
  • windows-aarch64

🛠️ Codebase & Quality Enhancements

  • Strict Java 21 Records: All data transfer objects and configurations (SubagentConfig, AgentResponse, UsageMetadata, etc.) leverage modern record semantics.
  • End-to-End Live Testing: Verified live streaming, tool calling, and local inference against both Google Gemini and LiteRT LM running Gemma 4.

📥 Getting Started

Add the dependency to your pom.xml:

<dependency>
    <groupId>io.github.glaforge</groupId>
    <artifactId>antigravity-sdk-wrapper</artifactId>
    <version>0.2.17</version>
</dependency>

For offline or air-gapped deployments containing the native binaries:

<dependency>
    <groupId>io.github.glaforge</groupId>
    <artifactId>antigravity-sdk-harness</artifactId>
    <version>0.2.17</version>
    <classifier>all</classifier>
</dependency>

📝 Commits & Contributors

Features

  • da0e1e8: achieve full parity with upstream Antigravity SDK v0.1.18 and LiteRT local inference

Tasks

  • 2634ba3: Releasing version v0.2.17
  • b03dae0: Bump for next development cycle

Special thanks to @glaforge and the Antigravity team!

Contributors

glaforge
Assets 2
Loading